Imagine if an AI model was tied to a set of rules, so tightly that changing one character in those rules would render the entire model useless. This isn’t just a thought experiment – it’s a real concept called the Silicon Accord, which uses cryptography to bind an AI model to a constitution.
So, how does it work? The process starts with training a model normally, which gives you a set of weights. Then, you hash the constitution text, which creates a unique code. This code is used to scramble the weights, making them useless without the original constitution.
When you want to run the model, it must first load the constitution, hash it, and use that hash to unscramble the weights. If the constitution is changed, even by one character, the hash will be different, and the weights will be scrambled in a way that makes them unusable.
This approach has some interesting implications. For one, it provides a level of transparency and accountability, since any changes to the constitution will be immediately apparent. It also means that the model is literally unable to function without the exact constitution it was bound to, which could be useful for ensuring that AI systems are used in a way that aligns with human values.
One potential challenge with this approach is that it requires a lot of computational power to unscramble the weights in real-time. However, the creators of the Silicon Accord have developed a solution to this problem, which involves keeping the weights scrambled even in GPU memory and unscrambling them just before each matrix multiplication.
Overall, the Silicon Accord is an innovative approach to ensuring that AI models are aligned with human values. By binding a model to a constitution using cryptography, we can create systems that are more transparent, accountable, and aligned with our goals.